Semantic Segmentation for outdoor scenes using simulated data

Abstract

In this work, we are studying how to do semantic segmentation for trees using simulated dataset. Semantic segmentation usually needs a big dataset to train a high performance model, but manually image annotation is very time consuming which is one of the biggest bottlenecks of the task. In this work we use a game simulator which provides methods for data collection without manual annotation, but is the data generated in the virtual world able to train a model that can understand things in real-world? Based on the question, we conduct our curiosity-driven research.